A new study from researchers including Mo Shakiba, Rana Rokni, Mohammad Mohammadi, and Nima Dehghani shows that building neural networks with real brain wiring and geometry helps them learn more efficiently. The team used data from the MICrONS program, which mapped over 12,000 excitatory neurons in mouse visual cortex by combining calcium imaging with electron microscopy. They used the neurons' spatial coordinates, anatomical connections, and functional relationships to initialize recurrent weights and impose spatial constraints during learning.
The Research
The scientists tested their biologically grounded recurrent neural networks on three cognitive decision-making tasks. The fully constrained networks outperformed baseline models and partially constrained versions. Functional weight initialization gave the biggest boost, but real spatial embedding also provided consistent improvements. These brain-like networks developed low-entropy, modular, small-world organization, similar to biological brains. Even when limited to positive weights, they retained strong performance. The paper was published on arXiv on June 12, 2026.
Why It Matters
This research shows that the cortex's geometry, wiring, and function are powerful inductive biases for building smarter algorithms. For your own brain, it highlights how structure and function are tightly linked—maintaining healthy neural connections through learning and mental activity may keep your cognitive networks efficient.
What You Can Do
Engage in activities that challenge multiple brain areas simultaneously, like learning a new language or playing strategy games. These tasks can strengthen functional connections and potentially keep your neural networks flexible.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
